Metadata-Version: 2.4
Name: ask-bedrock
Version: 0.3.2
Summary: Converse with your favorite Amazon Bedrock LLM from the command line.
Home-page: https://github.com/awslabs/ask-bedrock
Author-email: Ulrich Hinze <uhinze@amazon.de>, Alexander Günsche <lxg@amazon.de>
Keywords: ask-bedrock,GenAI,Bedrock,AWS
Classifier: Development Status :: 4 - Beta
Classifier: Topic :: Utilities
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
License-File: NOTICE
Requires-Dist: click
Requires-Dist: pyyaml
Requires-Dist: boto3==1.37.38
Requires-Dist: mcp
Provides-Extra: test
Requires-Dist: pytest>=7.0.0; extra == "test"
Requires-Dist: pytest-cov>=4.0.0; extra == "test"
Requires-Dist: mock>=5.0.0; extra == "test"
Dynamic: license-file

# Ask Amazon Bedrock

Converse with your favorite [Amazon Bedrock](https://aws.amazon.com/bedrock/) large language model from the command line.

<p>
  <img width="1000" src="https://raw.githubusercontent.com/awslabs/ask-bedrock/main/README.svg">
</p>

This tool is a wrapper around the low-level Amazon Bedrock APIs. Its main added value is that it locally persists AWS account and model configuration to enable quick and easy interaction.

## Installation

⚠️ Requires Python >= 3.9

⚠️ Requires a working AWS CLI setup configured with a profile that allows Amazon Bedrock access. See [CLI documentation](https://docs.aws.amazon.com/cli/latest/userguide/cli-configure-files.html) for details.


```shell
pip install ask-bedrock
```

Please refer to the troubleshooting section below if this does not work.

You can also build/run this project locally, see [Building and Running Locally](#building-and-running-locally).

## Usage

### Activating models

Before you can use this command line tool, you need to [request model access through the AWS Console](https://console.aws.amazon.com/bedrock/home#/modelaccess) in a region where Bedrock is available: Switch to the region where you want to run Bedrock, go to ”Model access“, click “Edit”, activate the models you wish to use, and then click “Save changes”.

### Invocation

To start a conversation, simply enter the following command:

```shell
ask-bedrock converse
```

If you don't need a conversation, you can get a simple request-response using:

```shell
ask-bedrock prompt "What's up?"
```

Upon the first run, you will be led through a configuration flow. To learn more about configuration options, see the [Configuration](#configuration) section below.

If you’re fully configured, the tool will show you a `>>>` prompt and you can start interacting with the configured model.

Multi-line prompts can be wrapped into `<<< >>>` blocks.

To end your interaction, hit `Ctrl + D`. Note that the conversation will be lost.

You can also use a single prompt with a simple request-response:
```
ask-bedrock prompt "complete this sentence: One small step for me"
```

### MCP
Ask Amazon Bedrock supports the [Model Context Protocol](https://modelcontextprotocol.io). You can register MCP servers through [configuration](#configuration), which auto-discovers resources and tools from an MCP server. The resources and tools are then available during invocation.

### Pricing

Note that using Ask Amazon Bedrock incurs AWS fees. For more information, see [Amazon Bedrock pricing](https://aws.amazon.com/bedrock/pricing/). Consider using a dedicated AWS account and [AWS Budgets](https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html) to control costs.

## Configuration

*Ask Amazon Bedrock* stores your user configuration in `$HOME/.config/ask-bedrock/config.yaml`. This file may contain several sets of configuration (contexts). For instance, you can use contexts to switch between different models. Use the `--context` parameter to select the context you'd like to use. The default context is `default`.

If no configuration is found for a selected context, a new one is created. If you want to change an existing config, use

```shell
ask-bedrock configure --context mycontext
```

You can also create or edit the configuration file yourself in `$HOME/.config/ask-bedrock/config.yaml`. Note that MCP configuration is verbose, but mostly auto-discovered during `ask-bedrock configure`:

```yaml
contexts:
  default:
    region: ""                  # an AWS region where you have activated Bedrock
    aws_profile: ""             # a profile from your ~/.aws/config file
    model_id: ""                # a Bedrock model, e.g. "ai21.j2-ultra-v1"
    inference_config: "{}"      # a JSON object with inference configuration
    mcp_servers:
    - command: npx
      args:
      - -y
      - '@modelcontextprotocol/server-filesystem'
      - /Users/uhinze/Downloads
      env: {}
      name: file
      resources: []
      tools:
      - description: Read the complete contents of a file from the file system. Handles various text encodings and provides detailed error messages if the file cannot be read. Use this tool when you need to examine the contents of a single file. Only works within allowed directories.
        inputSchema:
          $schema: http://json-schema.org/draft-07/schema#
        additionalProperties: false
        properties:
          path:
            type: string
        required:
        - path
        type: object
        name: read_file
        server_name: file
      - <...more tools>
```

### Inference Configuration

The `inference_config` is passed directly to the Amazon Bedrock Runtime `converse_stream` API. This configuration controls the behavior of model generation, including parameters like temperature and token limits.

Common parameters include:

- `temperature` (float): Controls randomness in response generation. Lower values make responses more deterministic.
- `topP` (float): Controls diversity of responses by considering tokens with top cumulative probability.
- `maxTokens` (integer): Maximum number of tokens to generate in the response.
- `stopSequences` (array): Sequences where the model should stop generating.

Example configurations:

```json
{
  "temperature": 0.7,
  "topP": 0.9,
  "maxTokens": 3000
}
```

```json
{
  "temperature": 0.5,
  "maxTokens": 500,
  "stopSequences": ["\n\n"]
}
```

For more details, see the [Amazon Bedrock Runtime InferenceConfiguration API Reference](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InferenceConfiguration.html).

## Building and Running Locally

```
pip install build
python -m build
pip install -e .
ask_bedrock converse
```


## Troubleshooting

**Q:** pip does not install ask-bedrock properly, instead it gives me an error like `This environment is externally managed`.

**A:** ask-bedrock needs to be installed globally to be available in your command line. Please follow the guidance in the error message by first installing `pipx` and then installing ask-bedrock using pipx.

---

**Q:** The model responses are cut off mid-sentence.

**A:** Configure the model to allow for longer response by increasing the `maxTokens` value in the inference configuration (see above). For example: `{"maxTokens": 3000}`

---

**Q**: I'm getting an error that is not listed here.

**A**: Use the `--debug` option to find out more about the error. If you cannot solve it, create an issue.

## Security

See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for more information.

## License

This project is licensed under the Apache-2.0 License.
